Improving classification-based diagnosis of batch processes through data selection and appropriate pretreatment

Abstract

a b s t r a c tThis work considers the application of classification algorithms for data-driven fault diagnosis of batch processes. A novel data selection methodology is proposed which enables online classification of detected disturbances without requiring the estimation of unknown (future) process behavior, as is the case in previously reported approaches.The proposed method is benchmarked in two case studies using the Pensim process model of Birol et al. (2002) implemented in RAYMOND. Both a simple k Nearest Neighbors (k-NN) and complex Least Squares Support Vector Machine (LS-SVM) are employed for classification to demonstrate the generic nature of the proposed approach. In addition, the influence of different data pretreatment methods on the classification performance is discussed, together with a motivation for selecting the correct pretreatment steps. Finally, the influence of the number of available training batches is studied.The results demonstrate that a good classification performance can be achieved with the proposed data selection method even with a low number of faulty training batches by exploiting knowledge on the nature of the to-be-diagnosed faults in the data pretreatment. This provides a proof of concept for classification-based batch diagnosis and demonstrates the importance of incorporating process insight in the construction of data-driven process monitoring and diagnosis tools.

An extensive reference dataset for fault detection and identification in batch processes

“…This enables unrealistically tight control of the process around its temperature and pH set points, greatly facilitating SPM. Many authors already recognized this problem and manually added measurement noise to the simulation data (e.g., [2][3][4]6,14,32,35,37,50,79,87,93]). In addition, Pensim only simulates a limited set of process upsets.…”

“…More recently, Gins et al [35] and Wuyts et al [85] also used Pensim to compare the performance of k Nearest Neighbors (k-NN) and Least Squares SVM (LS-SVM) classifier. They employed 200 normal and 6600 faulty batches (1100 each of 6 types), providing full specifications on the type, magnitude and onset time of the various faults.…”

Industrial Process Monitoring in the Big Data/Industry 4.0 Era: from Detection, to Diagnosis, to Prognosis

“…However, this apparently simpler methodology also raises some practical issues that need to be addressed. A first issue faced by simultaneous fault detection and identification is that a large data imbalance typically exists between the normal operation class and one or more fault classes, as faulty examples are difficult to obtain [123][124][125]. This presents a significant challenge for most multi-class classifier types [126][127][128][129].…”